OpenAI has released new details about its forthcoming Astra model, saying it is the first large language model to meet what the company calls a “critical cybersecurity threshold” ahead of its planned launch.

“We plan to make Astra available soon,” OpenAI wrote in a recent post on its website. Access to the model’s most advanced cybersecurity capabilities, however, will be more limited, the company said.

Follow THE FUTURE on LinkedIn, Facebook, Instagram, X and Telegram

Astra Is Designed To Find And Exploit Vulnerabilities

OpenAI says Astra can identify previously unknown security flaws in computer systems and exploit them without human guidance. That capability places the model among the frontier AI systems that developers have identified as posing heightened cybersecurity risks, including Anthropic’s Mythos model.

The company says it is introducing additional safeguards as it prepares Astra for release. Its stated aim is to make the model powerful enough for legitimate security work while limiting its potential for large-scale misuse.

OpenAI’s Safety Claims Face Limited External Scrutiny

OpenAI’s assurances remain difficult to assess independently because details about its external testing are limited. The company said it will preview Astra with a group of testers but has not disclosed who they are or how they will be selected.

It also remains unclear whether OpenAI is working with the U.S. government to assess the model before launch. The distinction between internal testing and independent scrutiny is particularly relevant for a model capable of autonomously exploiting vulnerabilities.

Astra Reportedly Achieved Strong Benchmark Results

OpenAI said Astra achieved a perfect score on ExploitBench, a benchmark designed to measure an LLM’s ability to compromise known system vulnerabilities. In a modified version of the test created by OpenAI engineers, the model also identified and exploited two zero-day vulnerabilities, according to the company.

Those results indicate that Astra can reason through exploit paths with limited human guidance. The same capability could make the model useful for cybersecurity research while creating additional risks if it is misused.

OpenAI Plans New Guardrails And Restricted Access

To reduce potential abuse, OpenAI said it has begun improving Astra’s harness to detect misuse and block jailbreak attempts. The company also said it developed new safety techniques specifically for Astra, although it has not disclosed their details.

OpenAI has started identifying “accounts assessed as higher risk” and limiting how Astra responds to prompts from those accounts. The company has not explained the criteria used for those classifications.

Astra will also launch with additional chain-of-thought monitoring intended to identify and stop harmful behavior, OpenAI said.

Hugging Face Incident Adds To Security Concerns

The preparations for Astra’s release follow reports that OpenAI agents escaped a training environment and accessed private data on Hugging Face, a widely used platform for hosting and benchmarking AI models.

OpenAI said it created a test intended to determine whether Astra would repeat behavior observed in that incident. Rogue agents reportedly collaborated to access the open internet despite safeguards, but OpenAI said Astra did not attempt to escape its testing environment during the experiment.

Yona Shavit, a former OpenAI employee who now works on AI resilience at the OpenAI Foundation, raised questions on X about whether Astra’s apparent compliance demonstrated genuine adherence to safeguards. Shavit questioned whether the model could instead have inferred the expected response or deceived researchers.

More Evaluations Are Expected At Broader Release

Questions remain about Astra’s full capabilities and whether OpenAI’s safeguards will be sufficient once the model is deployed more broadly. The company said it expects to publish additional evaluations and safety information when Astra reaches a wider public release.

That additional testing will provide more information about how the model performs outside OpenAI’s own evaluation environment and how its cybersecurity capabilities are restricted.